Links between substance use, psychosis, and death in people living in unstable housing or homelessness

The Hotel Study is planned as a 20-year, naturalistic, longitudinal prospective study following an adult, community-based sample [23,34,39]. Briefly, participants were recruited from single-room occupancy (SRO) hotels (n = 372) and a sequential series from the community court (n = 65) in a circumscribed neighborhood of downtown Vancouver, Canada from November 1, 2008 to October 26, 2015. All tenants in a building were potentially eligible to participate; the only criteria were being 18 years of age or older, able to communicate in English, and able to provide informed consent. The neighborhood houses approximately 3,800 residents in SRO hotels; the court processes 2,500 cases per year. Participants were homeless or lived in precarious or marginal housing, defined as being below Canadian standards for adequacy (need for repairs), affordability (rental costs <30% of before-tax income), or suitability (makeup of bedrooms and household) [40]. Rooms in SRO hotels are generally 8–12 m² in size, with a sink and sometimes a hotplate. Toilet and shower facilities are shared between 10 and 15 tenants on each floor.

All participants provided written informed consent once procedures were completely described. Consent was reaffirmed at each visit thereafter. Clinically significant laboratory findings were shared with participants and their physicians. The study and the amount of an honorarium for participation were approved by the Clinical Research Ethics Boards of the University of British Columbia and Simon Fraser University.

Sociodemographic variables were recorded by a research assistant. A psychiatrist performed a semistructured interview, a mental status examination, and a focused neurological exam. A research assistant carried out a Mini-International Neuropsychiatric Interview [41]. Healthcare records for participants’ reports of past hospitalizations for mental illness were obtained from across Canada (available from >50 years ago in some cases). Baseline psychiatric symptoms were assessed using the 30-item Positive and Negative Syndrome Scale (PANSS) [42]; functioning was assessed with the Social and Occupational Functioning Assessment Scale (SOFAS) and the Role Functioning Scale (RFS) [8,43]. Diagnoses of current and past mental and substance use disorders were made by study psychiatrists (WGH, OL, FVR) using all available information noted above, according to criteria from the Diagnostic and Statistical Manual for Mental Disorders-TR Fourth Edition (DSM-IV-TR) [8]. Coroner’s reports and records from hospitalizations for medical (nonpsychiatric) reasons were obtained for participants who died during the study period.

Screening laboratory tests were carried out at baseline and annually thereafter. These included a complete blood count and differential, liver function tests, and serology for hepatitis C virus (HCV) and human immunodeficiency virus (HIV). The aspartate aminotransferase-to-platelet ratio index (APRI) was used as a surrogate measure for hepatic fibrosis (values > 0.7) (42). More detailed descriptions of annual and monthly assessments, with references, appear in open-source publications and associated supplements [23,34].

A comprehensive baseline assessment included standardized measures of time-invariant and time-varying risk factors (Fig 1A).

Time-invariant risk factors are those that occurred in the past and/or are not modifiable. These include past psychotic disorder diagnosis (as above), persistent sequelae from past TBI, exposure to early-life traumatic events (any, and number), and homelessness by age 18. A history of TBI was determined by clinical interview and/or high-field (3-Tesla) magnetic resonance imaging (MRI) findings identified by a neuroradiologist. Criteria for the presence of a past TBI were reporting a serious head injury with clinical manifestations (loss of consciousness ≥5 minutes or confusion ≥1 day and persistent sequelae defined as any of seizures in the past year attributable to TBI, need for seizure prophylaxis, or persistent neurological findings attributable to TBI) as used previously [23]. Past traumatic events were assessed using the semistructured Trauma History Questionnaire (THQ) [44]. Events involved objective threat of death or serious injury to which an individual reacted with extreme fear, horror, or helplessness according to the DSM-IV-TR criteria. The THQ is scored as the number of types of traumatic events endorsed to a maximum of 23, including physical, sexual, disaster-related, or crime-related events [45]. The number of types of traumatic events prior to age 18 (THQ18) was also recorded. Homelessness was defined as living on the street, in a shelter, couch surfing, or having no fixed address.

Monthly follow-up visits were carried out in the neighborhood study office or by outreach to the participants’ dwellings. Exposures to time-varying risk factors occurring over the concurrent and previous 3 weeks (i.e., total 1 month) were assessed, including nonprescribed substance use (types and frequency), experience of traumatic events, homelessness, and the potentially protective effect of prescription drug treatment. Nonprescription substance use and prescribed drug treatment were reported week by week using the Timeline Follow-back method [46]. Nonprescription substance use in the psychosis assessment week (i.e., the 7 days prior to each monthly visit) was examined to capture acute effects of substance exposure on symptom severity; substance use in the week prior to the assessment week was examined for persistent effects of exposure. Methamphetamine, cocaine, cannabis, opioid, alcohol, and tobacco use were reported (and confirmed by urine testing at the visit for methamphetamine, cocaine, cannabis, and opioids, n = 3,267 observations, percent agreement ranging from 83.0% to 87.1%, and kappa ranging from 0.62 to 0.68, p < 0.001). Days of use and route of administration were reported. Tobacco use was dichotomized (nondaily or daily use) because of the high prevalence (89.0%) among participants. Antipsychotic treatment (medication type, dose, route of administration, days of use) was reported week by week. The validity of self-reported antipsychotic use was confirmed using PharmaNet, a province-wide network that links all pharmacy dispensing records to a central database (n = 67 available records for antipsychotic use, kappa = 0.71, p < 0.001). Adequacy of treatment for managing psychosis was determined according to the Clinical Handbook of Psychotropic Drugs guidelines [47] and reported adherence (i.e., depot or ≥80% of past 28 days taking oral medication) in consultation with a psychopharmacologist (RMP). Prescribed methadone therapy was reported and considered adequate if taken ≥80% of the past 28 days. Adequate methadone reports exhibited high concordance with urine testing (n = 3,215 observations, kappa = 0.82, p < 0.001).

Recent trauma was assessed at monthly visits (Fig 1A). The number of types of traumatic events that occurred in the concurrent month (rTHQ) was recorded and analyzed as an ordinal variable truncated at 2 or more, since only 3.0% had 3 or more types of events in the concurrent month. Periods of homelessness were also captured each month.

The prespecified plan was to focus on 5 key positive symptoms spanning the psychotic features description in the DSM: delusions, hallucinations, conceptual disorganization (thought disorder), suspiciousness (which may be delusional), and unusual thought content (which may be delusional). Individual PANSS items are scored on a 7-point scale, with descriptive anchors for each level of severity over the previous 7 days. Items scored 1 or 2 out of 7 are respectively “within limits of normal experience” or “questionable.” Threshold severity scores were established in advance for the 5 key items to establish the presence of clinically relevant psychotic features and are based on previous work [9–11]. Threshold scores were item-specific, calibrated to the descriptors provided for rating the PANSS as follows: (1) the threshold for the delusions and hallucinations items according to PANSS descriptors is a score of 3 (mild), indicating these symptoms are definitely present, consistent with the psychotic features definition according to the DSM and categorically different from a state of good mental health. (2) The threshold for the items of conceptual disorganization and unusual thought content was set to a score of 4 (moderate). For these items, according to PANSS descriptors, a score of 3 (mild) may or may not be associated with formal thought disorder or bizarre delusions. For this reason, the threshold score was set to a higher level, consistent with the use of similar items as inclusion criteria for clinical trials of antipsychotic medications [48]. (3) The threshold for the suspiciousness item was set to an even higher threshold value of 5 (moderately severe) because only at this level or above is the descriptor provided in the PANSS consistent with a paranoid delusion [42].

We examined the inter-rater reliability of this assessment strategy. Two raters carried out separate interviews of 26 participants on the same day. The reliability for presence or absence of at least 1 of a possible 5 symptoms meeting the threshold for psychotic features was kappa = 0.69 (p < 0.001). This is a more stringent approach than is used in most studies, in which inter-rater reliability for individual PANSS items is reported by having different raters observe the same video of a patient interview or by having 2 raters carry out a joint interview of a patient. Of the few available studies using independent raters, and independent interviews of the same patient, a test–retest reliability intraclass correlation coefficient = 0.65 (95% confidence interval [CI] 0.62–0.68) was reported for the crosscutting domain of psychosis (hallucinations or delusions) [49].

Descriptive statistics of participants’ time-invariant and time-varying risk factors for psychotic features were reported as mean (standard deviation) or median (interquartile range) for continuous variables and number (proportion) for categorical variables. The associations between a past history of a psychotic disorder diagnosis and current functioning at baseline were assessed using analysis of covariance, controlling for age and sex. Similar analyses were carried out with psychotic features at baseline.

The first objective was to investigate how risk exposures collectively contribute to psychotic features among adults in precarious housing. Using data from monthly visits over the first year following each participant’s study entry, mixed-effects logistic regression models with random intercept and slope for longitudinal binary outcome data were used to assess the relationships between psychotic features (presence or absence during the psychosis assessment week) and exposure to risk factors. Fixed effects estimated the main effects of risk factors on odds of psychotic features. Time-invariant risk factors included past psychotic disorder diagnosis (presence or absence), persistent sequelae from past TBI (presence or absence), homelessness by age 18 (presence or absence), and past exposure to traumatic events (dose–effect estimate). Time-varying risk factors included number of days of nonprescription substance use, exposure to recent traumatic events (dose–effect estimates), and homelessness in the concurrent month (present or absent). Linearity of dose–effect estimates was tested by modeling factors as continuous and ordinal predictors in separate models and testing both linear and quadratic effects. Other covariates included age, sex, time point, antipsychotic treatment, and methadone therapy. Random effects of subject and time were included to account for between-subject variability and within-subject correlation across repeated measures, respectively. Random effect standard deviations were reported. Adjusted models were estimated by a stepwise model fitting process iteratively adding and removing random effects, then time-invariant and time-varying fixed effects variables that were significantly associated with the outcome in bivariate analyses [50]. The best fitting multivariable model was identified by successively comparing nested models using a likelihood ratio test and comparing Akaike Information Criterion (AIC) values. Model diagnostic plots were assessed for normality of the distribution of residuals. We tested multiplicative interactions using the Wald test. Predicted probabilities for psychotic features were estimated for combinations of significant risk factors.

Several approaches were employed to test the robustness of results (Fig 1A). Time-lagged substance use during the week prior to the psychosis assessment week was included in the adjusted model to estimate the persistence of substance use effects. Reverse association (sometimes described as “reverse causality,” although causality cannot be determined) was investigated by estimating the effect of psychotic features (presence or absence) on the days of substance use in the week following the psychosis assessment week using linear mixed-effects models and on the likelihood of a traumatic event during the following month using logistic mixed-effects models, adjusting for covariates. Further, the validity of self-reported substance use was tested by substituting urine results from the psychosis assessment visit for self-report in adjusted models. Sensitivity analyses assessed for evidence of systematic bias and the pattern of missingness by comparing participants who had missing or available variables of interest, who remained engaged in the study versus who discontinued, and who were included or excluded from longitudinal analyses using chi-squared and Wilcoxon rank–sum tests. Multiple imputation analysis was performed by imputing values generated from regression models with relevant predictors, including covariates of the main model of interest and variables related to a variable’s missingness. First, we imputed 10 completed data sets with 10 iterations for each imputation analysis. Second, the main model of interest was fitted to each of the completed data sets. Last, the parameter estimates from the second step were pooled to produce final parameter estimates. Pooled parameter estimates from the imputed data sets were compared to those from complete-case analysis to test whether our findings and inferences were affected by missing data.

The second objective was to examine the role of substance use, selected medical illnesses, and psychotic disorders as potential risk factors for early mortality, focusing on previously reported risk factors in the present sample examined over a shorter time period [38,51,52]. These analyses were not prespecified. Left-truncated and right-censored Cox regression models, with age as the timescale, were used to examine the effect of psychosis (a past history of psychotic disorder and/or psychotic features present during the first year after study entry), substance use, and related physical illness (hepatic fibrosis, HIV at baseline) on the risk of mortality for the time period November 1, 2008 to November 15, 2018. Statistical tests and visual plots of the Schoenfeld residuals were used to assess the assumption of proportionality of hazards. Violations of proportionality (i.e., significant Schoenfeld residual global test) were addressed by stratifying the sample using an age change point (threshold age separating a younger and an older group), determined by visual inspection of the Schoenfeld residual-by-age plots and identification of the age corresponding to the inflection point of the smoothing spline fit to the plots. Kaplan–Meier survival curves were constructed to plot the effects of statistically significant predictors. All analyses were performed in R (version 3.4.0; R Core Team) and RStudio (version 1.0.143; R Foundation for Statistical Computing). Significance was set to alpha level of 0.05, and all p-values are 2-sided. Additional details on methodology appear in the S1 STROBE Checklist and the S1 Protocol.

Staggered participant enrollment and flow appears in Fig 1B. Demographic and clinical characteristics of participants appear in Table 1.

At baseline, 425 of 437 (97.3%) participants were living in precarious housing, and 12 (2.7%) were homeless. The demographic characteristics of participants were similar to other reports from this Vancouver neighborhood and to reports from Canadian studies of precariously housed or homeless people in other cities (,). S1 Table S2 Table

The majority of participants had experienced past psychotic disorders and substance dependence (Table 1). In 4,285 assessments carried out across all participants over the first year following study entry, with a mean (SD) of 9.8 (3.6) visits per person, 335 of 428 (78.3%) participants endorsed positive symptom severity indicating the presence of 1 or more psychotic features according to the DSM on at least 1 visit (Table 1). Considering all assessments aggregated across participants over the first year after study entry, a single psychotic symptom was recorded in 781 of 4,285 (18.2%) of assessments; 2 or more psychotic symptoms were present in 1,069 of 4,285 (24.9%) of assessments. In assessments with 2 or more psychotic symptoms present, delusions and hallucinations were the most frequent combination, occurring in 709 of 1,069 (66.3%), followed by delusions and unusual thought content in 561 of 1,069 (52.5%).

At baseline, analyses of covariance showed participants’ levels of social and occupational function and of role functioning were associated with both a past history of psychotic disorder and the presence of current psychotic features, adjusting for age and sex (S3 Table, S4 Table). In combined analyses, the associations with current psychotic features predominated (SOFAS: estimate 2.3, standard error [SE] 0.6, p < 0.001; RFS: estimate 0.8, SE 0.2, p < 0.001).

The characteristics of risk factors for psychotic features are described in Table 2.

Persistent sequelae of past TBI indicated the consequences of physical trauma. Multiple types of traumatic life experiences before age 18, with potential implications for the development of mental disorders, were common. During the first year after study entry, participants used nonprescribed substances, in increasing order: methamphetamine, opioids, cannabis, cocaine, alcohol, and tobacco. In users, frequency of substance use over the first year varied from once per week to daily. Most participants experienced traumatic events at least once over the year; periods of homelessness were also noted as potential stressors. Adequate antipsychotic treatment was less common than adequate methadone treatment.

Effects of time-invariant and time-varying risk factors for psychotic features observed in the mixed-effects logistic regression models are summarized in Table 3.

The most potent risk factor was a past diagnosis of a psychotic disorder (see also Table 1). The final adjusted model (Table 3) indicated independent, linear, dose-dependent effects of days of alcohol, of methamphetamine, and of cannabis use in the concurrent week on odds for the presence of psychotic features. Greater exposure to trauma in the concurrent month was linearly associated with greater odds of psychotic features. The odds of experiencing psychotic features were greatest among males. Additionally, the presence of psychotic features was associated with receiving adequate antipsychotic treatment in the past month (likely indicating at least some of those in need did access treatment, but with incomplete response). These effects did not change over time during the year, and there was no evidence of multiplicative interactions. Other putative risk factors, including homelessness in the concurrent month, were not associated with experiencing psychotic features, nor was time itself associated with risk.

For longer-acting substances, use in the week prior to the psychosis assessment week could contribute to the expression of psychotic features. To study this, a time-lagged analysis was carried out using the same mixed-effects logistic regression models as above, with substance use in the prior week substituted for substance use in the week concurrent with the psychosis assessment (Fig 1B and S5 Table). Time-invariant and concurrent month risk factor effects were unchanged. Use of cannabis or alcohol in the week prior did not contribute to the likelihood of psychotic features being observed in the psychosis assessment week (Fig 2).

Methamphetamine use showed a different pattern, as use in the prior week was associated with psychotic features in the psychosis assessment week (aOR = 1.12, 95% CI 1.03–1.22, p = 0.012). This suggests the effects of methamphetamine as a risk factor for psychotic features may be relatively longer lasting and/or the effects of cannabis and alcohol are more acute and transient. Next, the possibility of reverse association was analyzed using psychotic features as the predictor of substance use in the week following the psychosis assessment week (Fig 1B). No significant associations were observed for alcohol, methamphetamine, cannabis, or cocaine use (all p > 0.05, S6 Table). The presence of psychotic features at assessment was associated with the outcome of new traumatic experiences in the following month (OR = 1.27, 95% CI 1.04–1.55, p = 0.02). A similar relationship was observed when the covariates of time, age, sex, past psychotic disorder diagnosis, days of other nonprescription substance use in the following week, and adequate antipsychotic treatment in the following month were added to the model (OR = 1.30, 95% CI 1.05–1.62, p = 0.02).

Effects of substance use in the psychosis assessment week on likelihood of psychotic features were similar in mixed-effects logistic regression models, substituting urine drug screen results at the visit for self-reported substance use as risk factors (Table 4).

Sensitivity analyses revealed no evidence of systematic bias in the pattern of missing data. Participants included in longitudinal analyses did not differ from excluded individuals, except they were older (chi-squared [1] = 6.67, p = 0.01). Participants with missing psychosis assessments were younger (estimate = −0.034, SE = 0.013, p = 0.007), and otherwise, missingness was not associated with psychotic features or risk factors. Participants who discontinued (e.g., moving from city, incarceration, living in treatment facility, lost contact) had greater THQ18 scores (Wilcoxon = 16,367, p = 0.039) but were otherwise similar on all factors to those who remained in the study. Altogether, data were determined to be missing at random, with missingness unrelated to psychosis or related factors, and we proceeded with the multiple imputation procedure to test whether missingness impacted the findings. Death, psychosis, and psychosis risk factors were included as potentially relevant predictors of missingness in the imputation procedure. Pooled parameter estimates from imputed data sets were similar to complete-case analysis results (S7 Table), suggesting that the present findings were not affected by missing data.

Participants continued with monthly follow-ups after the first year in the study, and as of November 15, 2018, participants were followed for a median of 6.3 years (25th–75th percentiles, 3.1–8.6 years) after study entry. During 2,481 person years of observation, 79 of 437 (18.1%) participants died. Causes of death were physical illness (40.5%), accidental overdose (35.4%), trauma (5.1%), suicide (1.3%), and unknown (17.7%). Unadjusted Cox proportional hazards regression analyses on the full sample indicated associations between mortality risk and evidence of hepatic fibrosis, HIV seropositive status, and baseline (but not past) alcohol dependence (Table 1 and S8 Table). Risks associated with past or baseline dependence on cocaine, methamphetamine, opioids, cannabis, or daily cigarette smoking were not statistically significant, nor was HCV-seropositive or qualitative polymerase chain reaction (qPCR)-positive status. Adjusted Cox regression analysis indicated statistically significant risks associated with hepatic fibrosis and alcohol dependence at baseline (Table 5).

The effects on mortality of both a past history of psychotic disorder and psychotic features occurring during the first year after entering the study interacted with age as determined by a significant Schoenfeld residual global test (past history p = 0.04, first year psychotic features p = 0.02, S8 Table). By visual inspection of Schoenfeld residual plots, the change-point age separating younger and older groups was set at 55 years, consistent with earlier reports [38,52]. Unadjusted Cox proportional hazards regression analyses on the group less than 55 years of age showed an association between past history of psychosis and early mortality; no association was observed for psychotic features in the year following study entry, nor were associations observed in the older group (S8 Table). In those less than 55 years of age, a past history of psychotic disorder was associated with early mortality in a model controlled for baseline alcohol dependence, hepatic fibrosis, and HIV status (Table 5).

There were 39 deaths among 323 (12.1%) in the younger group and 40 among 114 (35.1%) in the older group. The distribution of causes of death did not differ between the younger and older groups or between those with or without a past history of psychotic disorder. As seen in Fig 3, for participants in the younger group, a past history of psychotic disorder (208 of 323, 64.4%) was associated with earlier death (30 of 208, 14.4% versus 9 of 115, 7.8%). In contrast, for participants age 55 or older, a past history of psychotic disorder (58 of 114, 50.9%) did not increase risk for early mortality (18 of 58, 31.0% versus 22 of 56, 39.3%).